Graph-Restricted Tensors

• Graph-restricted tensors are algebraic structures defined by a graph that constrains tensor entries and enforces specific entanglement patterns.
• They enable efficient tensor network models and precise control of correlations in quantum many-body physics, holography, and data analysis.
• These tensors facilitate advanced decompositions and signal processing techniques by integrating graph-based constraints to reduce computational complexity.

A graph-restricted tensor is an algebraic object whose entries and the constraints they satisfy are imposed by the combinatorial structure of an underlying graph. The framework serves as a unifying language for encoding entanglement/independence patterns, structural constraints, and symmetry properties in tensor network models, algebraic geometry, graph theory, quantum many-body physics, and applied data analysis. The graph restricts the allowable correlations, symmetries, or linear mappings between tensor indices, enabling precise control over entanglement, computational cost, and expressibility in multi-partite systems. Recent work demonstrates the power of this formalism in holographic tensor network models, tensor rank bounds, signal processing, quantum error correction, and combinatorial identification.

1. Formal Definition and Entanglement Constraints

Let $G=(V,E)$ be a simple undirected graph on %%%%1%%%% vertices $V$. A complex order-$n$ tensor $T_{i_1\cdots i_n}\in\mathbb{C}$ defines an unnormalized $n$-partite quantum state $$|\psi_T\rangle=\sum_{i_1\cdots i_n} T_{i_1\cdots i_n} |i_1\cdots i_n\rangle$$ in $(\mathbb{C}^d)^{\otimes n}$.

Graph-restricted tensor: $T$ is $G$-constrained if for every subset $C=\{v_{k_1},\ldots,v_{k_m}\}$ that forms a clique in $G$, the reduced density matrix $\rho_C$ is maximally mixed: $$(\rho_C)_{i_{k_1}\cdots i_{k_m}}^{j_{k_1}\cdots j_{k_m}} \propto \prod_{q=1}^m \delta_{i_{k_q}}^{j_{k_q}}.$$ Equivalently, for any bipartition with output indices corresponding to a clique, the map $V_T V_T^{\dagger}$ is proportional to identity, enforcing isometry/unitarity. If all maximally mixed reductions correspond to cliques, the tensor is faithfully $G$-constrained (Bistroń et al., 28 Dec 2025).

This clique-induced constraint generalizes multipartite entanglement patterns:

• Empty graph $\Rightarrow$ 1-uniform states (each qudit entangled with the rest)
• Two disjoint edges $\Rightarrow$ maximal entanglement across edge bipartitions
• Cycle/planar graph $\Rightarrow$ block-perfect/dual unitary tensors
• Complete graph $K_n$ $\Rightarrow$ absolutely maximally entangled (AME) states (perfect tensors, 2-unitary)

2. Graph-Restricted Tensor Rank and Complexity

Given a graph $G$, the canonical graph-restricted tensor is (Christandl et al., 2016, Christandl et al., 2016): $$T_n(G) = \sum_{i:E\to[n]} \bigotimes_{v\in V} \bigotimes_{e\ni v} b_{i(e)},$$ where $b_{i(e)}$ is the basis for edge labelings. $T_n(G)$ encodes all possible edge-assignments dictated by $G$.

• Tensor rank: $R(T)$ = minimum $r$ so $T$ is decomposed into $r$ simple tensors.
• Asymptotic rank: $$\widetilde{R}(T) = \lim_{n\to\infty} R(T^{\otimes n})^{1/n}$$; exponent $\omega(T) = \log_2 \widetilde{R}(T)$.
• Per-edge exponent: For $G$, $\tau(T(G)) = \omega(T(G)) / |E|$.

Main result: For complete graphs $K_k$, $\tau(T(K_k)) \leq 0.772943$ for $k\ge 4$, which is smaller than the best-known bound for matrix multiplication ($\sim 0.79$ for $k=3$). The threshold $2/3$ remains central: for $k=3,4$, lower bound equals $2/3$, and conjecturally $\tau(T(K_k)) \leq 2/3$ if matrix multiplication exponent $\omega=2$ (Christandl et al., 2016). Efficient tensor contraction and resource cost per edge in quantum protocols are governed by these exponents.

3. Tensor Network Varieties and Dimension Bounds

Graph-restricted tensor varieties parameterize all tensors expressible as contractions on a graph: $$V_{G, r} = \text{Zariski closure of } \{ (X_1 \otimes \cdots \otimes X_d) \cdot T(G, r) \mid X_v: W_v \to V_v \},$$ with $W_v=\bigotimes_{e\ni v} \mathbb{C}^{r_e}$ for edge bond dimensions $r_e$ and $V_v=\mathbb{C}^{n_v}$ for local dimension.

Dimension upper bound: $$\dim V_{G, r} \leq \min\left\{ \sum_{v=1}^d (n_v N_v) - d + 1 - \sum_{e\in E} (r_e^2 - 1), \prod_{v=1}^d n_v \right\}$$ where $N_v = \prod_{e\ni v} r_e$ (Bernardi et al., 2021).

When every vertex is strictly supercritical ($N_v \leq n_v$), this bound is sharp and the parameterization reduces to a single orbit under the gauge group; relevant for characterizing tensor dimensions in MPS, PEPS, and higher-dimensional varieties (see explicit bounds for cycles, grids).

4. Graph-Restricted Models in Signal Processing, Machine Learning, and Data

The concept extends beyond quantum and algebraic combinatorics to graph-regularized tensor decompositions, multi-way graph signal processing, and graphical latent-variable identification (III et al., 2020, Shahid et al., 2016, Sofuoglu et al., 2019, Gu, 18 Jan 2025).

Multilinear Low-Rank Frameworks

For a $d$-way tensor $\mathcal{Y}$, per-mode graphs $G_\mu=(V_\mu,E_\mu,W_\mu)$, spectral decomposition $L_\mu = P_\mu \Lambda_\mu P_\mu^T$ enables MLRTG decompositions: $$\mathrm{vec}(\mathcal{Y}) = (P_{1,k_1} \otimes \cdots \otimes P_{d,k_d})\,\mathrm{vec}(\mathcal{X}),\quad \mathcal{X}\in\mathbb{R}^{k_1\times\cdots\times k_d}$$ Smoothness and low nuclear norm in the graph-eigenbasis yield robust low-rank models for tensor PCA, completion, and RPCA; exact recovery governed by Laplacian eigen-gaps (Shahid et al., 2016).

Graph-Regularized TT/CP Decomposition

The GRTT model enforces both TT-dimensionality reduction and manifold preservation via Laplacian penalty: $$\min_{U_n,X} \|Y - \text{TT}(U_n, X)\|_F^2 + \lambda\, \mathrm{tr}(L(X)^T L L(X)),$$ with orthonormality (Stiefel) constraints, solved efficiently via ADMM (Sofuoglu et al., 2019). Empirically, GRTT achieves superior clustering accuracy, scalability, and storage efficiency compared to graph-regularized Tucker and CP models.

Tensor Unfolding for Graph Identifiability

In bipartite graphical models (e.g., RBM, Noisy-Or networks), graph-restricted tensor unfolding allows constructive graph recovery via rank signatures in population-level tensors:

• Rank concentrations in unfolded matrices reveal latent-variable connections.
• The identifiability condition: each latent connects to at least two pure observed nodes (Gu, 18 Jan 2025).

5. Graph-Restricted Tensors in Quantum and Holographic Networks

Graph-restricted tensors encode fine-tuned multipartite entanglement and operator constraints in quantum networks (Bistroń et al., 28 Dec 2025):

• The framework subsumes AME states, dual unitaries, perfect tensors, and non-stabilizer family constructions.
• Isometry and unitarity on cliques ensure solvable, analytically tractable holographic models—crucial for exactly computable correlation functions and scalable physical simulation.
• Exact solutions exist for a wide landscape of non-stabilizer graph-restricted tensors: e.g., hexagonal planar 7-qubit families and pentagonal AME(5,2) tensors.

Functionally, these tensors yield:

• Power-law decay of boundary correlation functions in holographic tilings
• Tunable scaling dimensions, central charge, and operator spectra in AdS/CFT toy codes
• Systematic generalization of HaPPY codes to imperfect yet tractable network components

6. Algebraic Connections: Homomorphism and Connection Tensors

In combinatorics and algebraic graph theory, homomorphism tensors and connection tensors offer a theory of graph-restricted functions (Grohe et al., 2021, Cai et al., 2019):

• Homomorphism tensor $\mathbf{F}_G$ encodes counts of labeled homomorphic images and is used to distinguish graphs up to isomorphism or within restricted classes (bounded treewidth/treedepth).
• Connection tensors $T(f, k, n)$ capture multiplicative graph parameters and generalize connection matrices to $n$-way arrays.

Exponential symmetric tensor rank bounds classify which graph parameters (partition functions) are expressible as vertex-and-edge-weighted homomorphisms: $T(f_H, k, n) \implies \rk_{\mathrm{sym}}(T(f_H, k, n)) \leq |V(H)|^k\;\forall k,n.$ Perfect matchings and Holant problems violate such bounds, demonstrating strict inexpressibility within the homomorphism partition framework—even over complex weights.

7. Generalizations and Future Directions

Graph-restricted tensor notions naturally extend to:

• Hypergraphs: encoding more intricate interaction patterns (higher-arity entanglement, PEPS/MERA networks)
• Multi-layer, heterogeneous network indices (hetero-functional graphs in MBSE) (Farid et al., 2021)
• Algebraic complexity, communication complexity, and quantum protocol resource theory (Christandl, 2023)

Open problems include efficient computation of graph-restricted ranks, finer analysis of solution varieties in large graphs, extension to directed/hypergraph settings, and algorithmic applications in model selection and isomorphism testing.

Summary Table: Main Classes of Graph-Restricted Tensors

Class/Model	Graph Constraint Type	Key Property/Use
Clique-constrained tensor	Entanglement/Isometry	Maximal mixedness on cliques; quantum codes
Homomorphism tensor	Image counts	Logical/structural equivalence; graph testing
Connection tensor	Partition functions	Algebraic classification of expressibility
MLRTG/GRTT/Decomp	Modewise Laplacians	Data compression; manifold/structure learning
Hetero-functional tensor	Multi-modal index sets	Multi-layer system ontology; MBSE analysis

The graph-restricted tensor paradigm provides a unified, rigorous, and highly versatile set of tools for encoding, analyzing, and utilizing constrained multilinear structure across mathematics, physics, computer science, and engineering.